VERA: Scaling Verifiable Environments for Agentic co-Evolution
paperYour notes
Long-horizon agent training is bottlenecked on environments, not on algorithms: an agent doing medical research "must ground a finding, classify it, and write a report over dozens of dependent steps, yet recent environments score only the outcome." VERA manufactures the missing environments and then trains on them. Each one is built from an initial trajectory — an agent writes rubrics and executable checks, a judge verifies the sandbox, and only those that pass enter the training bank, so the corpus is filtered by construction rather than curated afterwards. The released corpus is 9,000+ long-horizon verifiable environments.
The co-evolution is the claim. Training alternates between two kinds of update: train the model with rubric rewards, or edit the harness skills — and a verifier gates both model checkpoints and harness edits against explicit development-set acceptance criteria. The authors argue this attribution is what separates it from single-axis baselines, because "its updates target not only the cause but the outcome": when something improves, the gate records which axis moved it. That is a third position this fortnight on the model-versus-harness question that Model–Harness Interactions tracks — here, rather than fixing the harness or dissolving it into the model, both are trained and the attribution is measured.
At 9B, the model paired with its co-evolved agent beats the strongest baseline by 10.3 and 13.0 points across the two domains. At 27B it reaches 71.6 on AutoCoWorkBench and 80.7 on AutoMedBench, and the paper reports transfer to unseen workflows with general capabilities retained — the two checks a method that trains on synthesised environments most needs to pass, since the obvious failure is a model that learns the generator.
Eight of the seventeen authors are at NVIDIA, including Daguang Xu and Yucheng Tang, with UC Santa Cruz, UIUC, NUS, Tsinghua, Xi'an Jiaotong and Penn alongside. Code is Apache-2.0; the paper is CC BY-NC-SA 4.0. No model checkpoints on HuggingFace at filing, so the release is the environments and the framework rather than the agents.